Safety Score

[dskillz]

I assume you were not using Autosteer or TACC? (Though the answer would not help really definitively answer the question about masking of these events.)

I was not.

 

[fiveten]

Looks like "hard braking" is weighted more heavily than "aggressive turning" - see the Safety Score Indicator to play "what if" games.

The former is defined as "proportion of braking time spent braking with extreme force". If making many easy turns in your neighborhood reduces that score, I'll bet the same may be true for braking as well. Do you recall those before/after stats?

I think it only counts something as hard braking if you are going faster than 50 or something

edit: nvm, the 50 mph applies to unsafe following distance. I didn’t pay attention to what my hard braking was before

 

[Daniel in SD]

Why would it be more predictive of collision? According to that theory, it would be unsafe to drive down a steep hill at a constant velocity and safer to accelerate.

Because the model is just a curve fit. One reason that it might fit better could be that people who drive in hilly areas have more collisions.
Between 0.1g-0.3g of longitudinal force will always help your safety score, there are very few hills steep enough that slowing down isn't optimal.

 

[VT_EE]

Find drivers with a clean driving record and no accidents. Having no accidents is your best indication of a safe driver because it means you have a driver who drives safely and is also defensive in trying to avoid potential accidents.

Not that I disagree with everything they are checking. My ideal safety check algorithm would include:

• Insurance claim history for accidents (This is the single most important criteria)
• No detected red light running. (not currently checked)
• No detected stop sign running (5mph threshold accepted)
• No tailgating (already done and I agree 100%)
• Signal for lane changes and turning (shows someone is courteous and methodical with their driving. Not currently checked)

I understand what they are trying todo. I really do and I support the concept but NOT the current execution.

All I know is that this has completely ruined my enjoyment of the car. It's a Sunday morning and I would normally go on a nice drive but they've made me driving may car into some miserable BS gamified nonsense that actually has nothing to do with me being able to safely monitor FSD Beta nor safely drive my car.

Only AT-FAULT insurance claims should count. I'd be pissed if I got dinged because some dude rolled into me while I was stopped at a stop sign (which happened).

 

[Twiglett]

I'm sure something is busted in the safety beta for FCW.
Took a drive this morning and checked the score, it rated at 75 with a red FCW rating. Car is set to medium FCW and no alerts sounded.
So I did it again and took the same route, luckily no cars in front but still got a bad FCW rating, again no alerts.
The route takes me down a two lane route that gently weaves left, right, left, right with a slight elevation change that freaks AP out.
Not looking good for a decent rating when this is one or two main routes I take to get out of my area.

 

[J03L]

Is it also measuring your Safety Score when using Model 3 Performance's Track Mode?

 

[EVNow]

They definitely checked it for accuracy, how else would you calculate the coefficients?

LOL. You are using "accuracy" the way some devs do.

Accuracy in terms of how good the events collected are, do they really represent useful safety events, should we filter out some obviously spurious things. Think like a data scientist who wants to find the right signals and data to use - rather than a dev just running large amount of data through ML. Most important (and sometimes laborious) aspects that gets overlooked a lot is cleansing of data.

BTW, "good driver" for the purposes of insurance might be somewhat different from FSD Beta. For eg. a driver who needs to maneuver the car is various ways with very little margins to fit in a garage should actually be penalized because they are more likely to get into property damage scenarios. So, they may even need different kind of cleansing for FSD beta safety rating compared to insurance.

 

[fiveten]

I’m gonna find a long stretch or road and start and stop a lot to see if I can drop my 1.4% aggressive breaking over 149 miles so far (the vast majority of those miles being autopilot)

 

[AlanSubie4Life]

I was not.

Definitely seems like using AP to mask unsafe following is the only option in LA freeway traffic. Seems like that should help clean up the scores and take you to over 99. Just have to avoid those hard braking events off the freeways (so far so good for you, but surprisingly difficult actually - though getting to 1-2% should not be too difficult). You really can't use the brakes or if you do you have to be very judicious about not creating a transient deceleration spike before obtaining the right level of stopping force (ease into the brakes). This is tough if you're stopping for a yellow light, for example. That initial brake "bite" might be what triggers a few % of hard braking; it's just kind of natural to have an initial deceleration pulse as you suddenly transition from accelerator to brake in response to an event.

if you reboot your system that trip won’t be recorded in your safety score calculation!

Let the gaming begin! Hopefully Tesla will patch this bug.

 

[VT_EE]

The "hard braking" is the one metric I find that makes this gaming unsafe. I rolled through a stale yellow today for fear of causing a score ding (which it would have). Normally I would have played it safe and stopped. Almost any pressure on the brake pedal will cause a score ding.

 

[DCEV]

Only AT-FAULT insurance claims should count. I'd be pissed if I got dinged because some dude rolled into me while I was stopped at a stop sign (which happened).

Fair point.

I wish there was a process for the FSD Beta access that too into account your driving record and insurance claim history for at fault accidents as THAT actually would be an indication of how safe and responsible you are.

 

[AlanSubie4Life]

The "hard braking" is the one metric I find that makes this gaming unsafe. I rolled through a stale yellow today for fear of causing a score ding (which it would have). Normally I would have played it safe and stopped. Almost any pressure on the brake pedal will cause a score ding.

Yeah, I guess Tesla wants us to use AP in those situations and just let the car figure out if it should blow the yellow or not. This is assuming that actually masks braking events (still sort of an open question, though it appears it should mask them according to the verbiage in the app which gives no caveats about appropriate use).

(Using AP in these situations will also make this gaming unsafe, in my opinion.)

 

[Daniel in SD]

LOL. You are using "accuracy" the way some devs do.

Accuracy in terms of how good the events collected are, do they really represent useful safety events, should we filter out some obviously spurious things. Think like a data scientist who wants to find the right signals and data to use - rather than a dev just running large amount of data through ML. Most important (and sometimes laborious) aspects that gets overlooked a lot is cleansing of data.

I'm saying they probably tried a bunch of models, fit them to that data, and measured how accurately they predicted collision rate. Using the proportion of >0.3g to 0.1g-0.3g braking was probably found to work better than a bunch of other possibilities.
Is there some other way to predict collision rate other than trying different models and seeing how well they work?

 

[andy.connor.e]

If folks are not seeing the Safety Score in the app, try logging out and back in
(not restarting the app: I mean sign off and re-enter your Tesla credentials)

it was because my app wasn't the correct version. installed through apk and I can see it.

 

[fiveten]

I’m gonna find a long stretch or road and start and stop a lot to see if I can drop my 1.4% aggressive breaking over 149 miles so far (the vast majority of those miles being autopilot)

Dropped it to 1.1% with a short drive just accelerating up to 15 mph and then decelerating to under 5 mph over and over again

 

[SarahsDad]

Looks like we get a score for each day of driving. So if you get a low score early in the day you can go out and drive more error-free miles and the score for the day increases. Don’t know if they just average the daily scores to get your weekly safety score but I’d guess yes. So you might be able to salvage a bad daily score by adding more miles that day.

 

[EVNow]

Is there some other way to predict collision rate other than trying different models and seeing how well they work?

Use ML software to create the model ?

But that should always come after cleansing the data. Use common sense / SME help to develop the rules to cleanse data.

 

[EVNow]

Don’t know if they just average the daily scores to get your weekly safety score but I’d guess yes. So you might be able to salvage a bad daily score by adding more miles that day.

Fairly sure, it would be calculated based on total number of infractions / total miles. Just averaging daily scores would give you wrong numbers.

 

[Knightshade]

On my 4th trip today got a "Unsafe Following" which makes no sense if UF only kicks in at 50 mph. I was using AP anytime I was driving 50 mph or higher.

Was this with AP turned on someplace Tesla says it shouldn't be used like a non-divided non-limited access highway?

If so it'd confirm that they ding you for APs bad behavior when using AP someplace they say it's not intended for.

 

[Daniel in SD]

Use ML software to create the model ?

But that should always come after cleansing the data. Use common sense / SME help to develop the rules to cleanse data.

They may already be cleansing the data. I also suspect that a ML model would be much better but then it would be a black box. Of course that might not be a bad thing since people changing their driving to score well on the model negates the stated purpose of the model.